What exactly is a Dirichlet process?

We start from the Dirichlet Distribution, which provides a vector of probabilities over states. The vector of probabilities must sum to 1, and they give Categorical Distribution probabilities, i.e. the probability of getting a draw from that category.

In contrast to the Dirichlet distribution, which has a fixed number of categories, the Dirichlet process describes a generative process for an infinite number of categories. There are a few algorithmic variants of the Dirichlet process, including the Chinese Restaurant Process, the Indian Buffet Process, and the Stick Breaking Process.

In practice, when computing with Dirichlet processes, we tend to use the Stick Breaking Process with a large but finite number of states.